Forcefield: A Local-First Go Runtime for AI Agents

Summary

Forcefield is a local-first command-line tool and single-binary runtime for running specialized AI agents in Go. It provides the operational layer around a model, including tools, skills, sessions, memory, permissions, context management, shell execution, project search, secret redaction, and communication with model providers. The runtime can connect to local models through Ollama, LM Studio, or llama.cpp, or to supported remote providers; a local model is optional, and the tool does not require a user account, cloud service, remote data processing, or telemetry. It offers an interactive terminal interface with streaming responses, direct task execution, session resumption, diagnostics, and slash commands for inspecting tools, skills, status, and sessions. Its built-in tools include file reading and writing, file and project search, shell commands, secret scanning, skill loading, and memory operations, with permission rules using allow, ask, or deny states. Shell execution can use the host directly or, on Windows, WSL with a pinned working directory and optional network isolation, although WSL does not confine filesystem access and Forcefield does not silently fall back to native execution. Sessions are stored per project, persistent memory is scoped per project, and session state is written atomically; long sessions use compaction and a provider-side sliding context window. Configuration lives in ~/.forcefield/config.yaml, while skills are Markdown files stored under ~/.forcefield/skills/. Linux, macOS, and Windows installers support amd64 and arm64 binaries, version pinning, upgrades, and checksum verification, though piping installation scripts directly from the main branch requires trusting their contents at execution time. The project is under active development, requires a supported model provider, and is licensed under Apache 2.0.